Filtered clips——

Pooled grapheme error—S + D + I / N

Pooled word / mixed—script-aware

Exact agreement—normalized text

Median SI-SDR—filtered selection

Median speaker similarity—ECAPA cosine

Language profileGrapheme errorSpeaker similarity

Error distributionClip count by GER

Metric ↔ textSI-SDRSpeaker similaritySTOILog-STFTMOS loss

Ranked results—

SortGrapheme error ↓Word / mixed ↓Speaker similarity ↑SI-SDR ↑MOS loss ↓Duration ↓Language A–Z

#	Clip ID	Lang	Duration	Grapheme	Word / mixed	Speaker	SI-SDR	Exact

⇤←—→⇥